CTL.SC1x -Supply Chain & Logistics Fundamentals. Forecasting for Special Cases

Transcription

1 CTL.SC1x -Supply Chain & Logistics Fundamentals Forecasting for Special Cases

2 Agenda Causal Analysis New Product Introductions Intermittent Demand Forecasting Wrap-Up 2

3 Causal Analysis 3

4 Causal Models Used when demand is correlated with some known and measurable environmental factor. Demand (y) is a function of some variables (x 1, x 2,... x k ) Dependent Variable Independent Variables Disposable Diapers ~ f(births, household income) Car Repair Parts ~ f(weather/snow) Promoted Item ~f(discount, placement, advertisements) 4

5 Example: Simple Linear Regression Recall from earlier lecture on exponential smoothing Estimating initial parameters for Holt-Winter (level, trend, seasonality) Removed seasonality in order to estimate initial level and trend Deseasoned Daily Bagel Demand y = x Time Period (Days) y i 0 1 x i Y i 0 1 x i i for i 1, 2,...n Observed Unknown EY ( x) x StdDev( Y x) 0 1 5

6 Simple Linear Regression The relationship is described in terms of a linear model The data (x i, y i ) are the observed pairs from which we try to estimate the Beta coefficients to find the best fit The error term, ε, is the unaccounted or unexplained portion The error terms are assumed to be iid ~N(0,σ) Deseasoned Daily Bagel Demand y = x Observed demand for period 97 = y 97 = 204 Error (residual) for period 97 = ε 97 = y 97 -ŷ 97 = = 15.6 Estimated demand for period 97 = ŷ 97 = (97) Time Period (Days) 6

7 Example: Monthly Iced Coffee Sales Develop Forecasting Model #1 Level and trend Develop OLS regression model Y i 0 1 x 1i i DEMAND = LEVEL + TREND(period) demand β 1 β 0 time 7

8 Model 1: Demand= b 0 + b 1 (time) R adj R s e RSS 917,500 TSS 4,093,071 Estimation Model Demand = (time) Coefficient Std Error (s bi ) Lower CI (95%) Upper CI (95%) t-stat p-value Intercept (b 0 ) 3, ,061 3, < Time (b 1 ) < Interpretation: Model explains ~77% of total variation in demand (very good!) Both the b 0 and b 1 terms make sense in terms of magnitude and sign and are statistically valid (p<0.0001) Captures level and trend nicely note how all data is used 8

9 Example: Monthly Iced Coffee Sales Develop Forecasting Model #2 Level, trend, & avg. historical temperature Develop OLS regression model Y i 0 1 x 1i 2 x 2i i DEMAND = LEVEL + TREND(period) + TEMP_EFFECT(temp) demand β 1 β 0 time 9

10 Model 2: Demand= b 0 + b 1 (time) +b 2 (temp) R adj R s e RSS TSS 917,074 4,093,071 Coefficient Std Error (s bi ) Lower CI (95%) Upper CI (95%) t-stat p-value Intercept (b 0 ) 3, ,891 3, < Time (b 1 ) < Temp (b 2 ) Interpretation: Adding temperature does not really have an effect on model fit Temperature is not statistically significant 10

11 Example: Monthly Iced Coffee Sales Other Forecasting Models Level, trend, & seasonality X Wi =1 if Mon=Dec, Jan, Feb; = 0 o.w. Y i 0 1 x 1i W x Wi i Level, trend, seasonality & school X Si =1 if school is in session, = 0 o.w. Y i 0 1 x 1i W x Wi S x Si i Many different ways to include different causal factors into a forecasting model. 11

12 Key Points Regression finds correlations between A single dependent variable (y) One or more independent variables (x 1, x 2, ) Coefficients are estimated by minimizing the sum of the squares of the errors Always test your model: Goodness of fit (R 2 ) Statistical significance of coefficients (p value) Some Warnings: Correlation is not causation Avoid over fitting of data Why not use this instead of exponential smoothing? All data treated the same Amount of data required to store Y i 0 1 x 1i 2 x 2i i 12

13 What are new products anyway? 13

14 New-to-World first of their kind, creates new market, radically different New-to-Company new market/category for the company, but not to the marketplace Line Extensions incremental innovations added to complement existing product lines and targeted to the current market 14

15 Product Improvements new, improved versions of existing offering, targeted to the current market replaces existing products Product Repositioning taking existing products/services to new markets or applying them to a new purpose Cost Reductions reduced price versions of the product for the existing market 15

16 Why does it matter? 16

17 New Product Categories Type of New Product Percent of Introductions Forecast Accuracy (1-MAPE) Launch Cycle Length Success Rate New-to-World New-to-Company 8-10% 17-20% 40% 47% 104 weeks 38-65% Line Extensions Product Improvements 21-26% 26-36% 54-62% 65% 62/29* weeks 55-77% Product Re-Positioning Cost Reductions 5-7% 10-11% 54-65% 72% N/A 66-79% * Major revisions / Incremental improvements about evenly split Adapted from Cooper, Robert (2001) Winning at New Products, Kahn, Kenneth (2006) New Product forecasting, and PDMA (2004) New Product Development Report. 17

18 Product-Market Matrix Product Technology Current Market Penetration New Product Development Current Forecasting Approach: Quantitative analysis of similar situations with item: time series, regression, etc. Forecasting Approach: Analysis of similar items: looks-like analysis or analogous forecasting Market Cost Reductions & Product Improvements Market Development Line Extensions Diversification New Forecasting Approach: Customer and market analysis to understand market dynamics and drivers Forecasting Approach: Scenario planning & analysis to understand key uncertainties & factors Product Repositionings New-to-Company & New-to-World Adapted from Kahn, Kenneth (2006) New Product Forecasting. 18

19 Why do firms launch new products? Companies earn significant revenue & profit from new products: Revenue - 21% to 48% Profit 21% to 49% By Selected Industries (revenue/profit): Fast Moving Consumer Goods 24% 24% Consumer Services 25% 24% Chemicals 18% 22% Healthcare 31% 33% Technology 47% 44% Product lifecycle is shortening and/or ability to maintain pricing is eroding faster Adapted from Cooper, Robert (2001) Winning at New Products, Kahn, Kenneth (2006) New Product forecasting, and PDMA (2004) New Product Development Report. 19

20 New Product Development Process 20

21 New Product Development Process Gate 1 Gate 2 Gate 3 Gate 4 Gate 5 Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 Stage 6 Discovery & Idea Generation Scoping & Pre-Technical Evaluation Build Business Case Development Test & Validate Commercialize Forecast Market Revenue Potential Forecast Market Revenue Potential Forecast Firm Sales Revenue Potential Forecast Unit Sales Forecast Unit Sales Forecast Unit Sales By Location 21

22 New Product Forecasting Methods Customer/market research 57% Jury of executive opinion 44% Sales force composite 39% Looks-like analysis 30% Trend line analysis 19% Moving average 15% Scenario analysis 14% Exponential smoothing 10% Experience curves 10% Delphi method 8% Linear Regression 7% Decision trees 5% Simulation 4% Others: 9% Methods differ by stage and by new product type. On an average, companies use 3 different methods to forecast new products. Business-to-Business (B2B) firms tend to use qualitative forecasts more than the Business-to-Consumer (B2C) firms. B2B firms have a longer forecasting horizon (34 months) compared to the B2C firms (18 months.) Adapted from Kahn, Kenneth (2006) New Product Forecasting. * Based on a survey of 168 companies.

23 Looks-Like or Analogous Forecasting How to do it Look for comparable product launches Create month by month (week by week) sales record Use the percent of total sales as guide to trajectory Similar to using comps in real estate Structured Analogy Create database of past launches (sales over time) Characterize each launch by Product type Season of introduction Price Target market demographics Physical characteristics Use characteristics of new product to query old launches Avoid only including successful launches! Adapted from Gilliland, Michael (2010) Business Forecasting Deal. 23

24 Forecasting Products with Intermittent Demand 24

25 Problems with Intermittent Demand Suppose I had an item that is ordered every 6 months for 1,000 units. Average monthly demand = 2000/12 = 167 units. What happens if I forecast demand using simple exponential smoothing: ˆx t,t 1 x t 1 ˆx t 1,t 0 1 Demand (units) Actual Demand x^(t,t+1) 25

26 Problems with Intermittent Demand Or more realistically, my demand for product is infrequent, of different size, and irregularly ordered. Forecasting demand using simple exponential smoothing results in additional noise. Separate components of demand and model separately: Time between transactions Magnitude of individual transactions Demand (units) Actual Demand x^(t,t+1) 26

27 Croston s Method Demand process: x t y t z t If demand is independent between time periods, then the probability that a transaction occurs is 1/n, that is: Prob y t 1 1 Prob y n t n The updating procedure becomes: If x t =0 (no transaction occurs), z^t = z^t-1 n^t = n^t-1 If x t >0 (transaction occurs), Forecast z^t =αx t + (1-α) z^t-1 n^t =βn t + (1-β) n^t-1 x^t,t+1 = z^t /n^t Where: x t = Demand in period t y t = 1 if transaction occurs in period t, =0 otherwise z t = Size (magnitude) of transaction in time t n t = Number of periods since last transaction α= Smoothing parameter for magnitude β = Smoothing parameter for transaction frequency Approach adapted from Silver, Pyke, & Peterson (1998), Inventory Management and Production Planning and Scheduling 27

28 Croston s Method An Example Using same data: Average demand per year ~2,000 units Irregular transaction size and time between orders Create a forecast for demand going forward using Croston s method =E7/F7 =IF(B6>0,1,D6+1) =IF(B7>0,$C$1*B7+(1-$C$1)*E6,E6) =IF(B7>0,$C$2*D7+(1-$C$2)*F6,F6) Demand (units) Actual Demand x^(t,t+1) 28

29 Croston s Method Essentially shifts the updating to only after an order occurs. Smooths out the forecast for replenishment purposes average usage per period Unbiased and has lower variance than simple smoothing. Cautions Infrequent updating introduces a lag to responding to magnitude changes Recommended use of smoothing for MSE of non-zero transaction periods NewMSE z x t ẑ t OldMSE z x t 0 29

30 Forecasting Wrap Up 30

31 Demand Process Three Key Questions What should we do to shape and create demand for our product? Demand Planning Product & Packaging Promotions Pricing Place What should we expect demand to be given the demand plan in place? Demand Forecasting Strategic, Tactical, Operational Considers internal & external factors Baseline, unbiased, & unconstrained How do we prepare for and act on demand when it materializes? Demand Management Balances demand & supply Sales & Operations Planning (S&OP) Bridges both sides of a firm Material adapted from Lapide, L. (2006) Course Notes, ESD.260 Logistics Systems. 31

32 Many Forecast Methods & Approaches Subjective Approaches Judgmental someone somewhere knows Experimental sample local and extrapolate Objective Approaches Time Series pattern matching Simple models (Moving Average, Cumulative, Naïve) Exponential smoothing - balancing new & old information Smoothing constants determine nervousness & response Lots of bookkeeping, updating & tricky initialization Causal Analysis underlying drivers Ordinary Least Squares (OLS) Regression A single dependent variable (y) and one or more independent variables (x1, x2, ) Testing the model and individual coefficients Watch-Outs: Correlation Causation & Avoid over-fitting Most firms use a portfolio of different techniques & methods a F b time time time time 32

33 Special Cases for Forecasting New Products No history there fore needed new methods Looks Like Forecasting Diffusion Models (innovator and imitator) Different types of new products Different methods New-to-World New-to-Company Line Extensions Product Improvements Product Re-Positioning Cost Reductions New product development process (stages & gates) Intermittent Demand Croston s Method smooths out sporadic and irregular transactions 33

34 Final Forecasting Comments Data Issues Dominate Sales data is not demand data Transactions can aggregate and skew actual demand Historical data might not exist Practical Things to Look For Forecasting vs. Inventory Management (avoid bias) Statistical Validity vs. Use and Cost of Model Demand is not always exogenous Error trending over time is it creeping? Hidden Costs of Complexity the more complex the system: the less frequently the parameters are checked and updated the less likely anyone who uses the system understands it the less likely operational teams will trust the output 34

35 CTL.SC1x -Supply Chain & Logistics Fundamentals Questions, Comments, Suggestions? Use the Discussion! Dutchess Photo courtesy Yankee Golden Retriever Rescue (

36 Image Sources 1 "Pampers packages ( )" by Elizabeth from Lansing, MI, USA - Pampers packagesuploaded by Dolovis. Licensed under Creative Commons Attribution 2.0 via Wikimedia Commons - "Kraft Coupon" by Julie & Heidi from West Linn & Gillette, USA - Grocery Coupons - Tearpad shelf display of coupons for Kraft Macaroni & Cheese. Licensed under Creative Commons Attribution-Share Alike 2.0 via Wikimedia Commons - "Damaged car door" by Garitzko - Own work. Licensed under Public domain via Wikimedia Commons - By Esa Sorjonen (I took the picture of my own Sony Walkman WM-2) [Attribution], via Wikimedia Commonshttp://commons.wikimedia.org/wiki/File%3ASony_Walkman_WM-2.jpg "PostItNotePad" by DangApricot (Erik Breedon) - Own work. Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons - "IPhone 2G PSD Mock" by Justin14 - Own work. Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons - "Ipod 1G" by Original uploader was Rjcflyer@aol.com at en.wikipedia - Transferred from en.wikipedia; transferred to Commons by User:Addihockey10 using CommonsHelper.. Licensed under Creative Commons Attribution 2.5 via Wikimedia Commons - "Ipod 2G" by Original uploader was Rjcflyer@aol.com at en.wikipedia - Transferred from en.wikipedia; transferred to Commons by User:Addihockey10 using CommonsHelper.. Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons - "Ipod backlight transparent". Licensed under Creative Commons Attribution-Share Alike 2.0-at via Wikimedia Commons - "Ipod 5th Generation white rotated". Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons - "IPod classic" by Kyro - Own work. Licensed under Creative Commons Attribution 3.0 via Wikimedia Commons - "IPod shuffle 1G" by Kyro - own work, about 2h with Adobe Photoshop CS4.. Licensed under Creative Commons Attribution 3.0 via Wikimedia Commons - "IPodphoto4G 1" by Original Photograph - AquaStreakImage Cleanup - Rugby471 - From English Wikipedia, original image is/was here. Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons

37 Image Sources 2 "Diet coke can" by The logo may be obtained from The Coca-Cola Company.. Licensed under Fair use of copyrighted material in the context of Diet Coke via Wikipedia - "Caff Free Diet cke can" by The logo may be obtained from The Coca-Cola Company.. Licensed under Fair use of copyrighted material in the context of Diet Coke via Wikipedia - "Vanilla coke zero can" by The logo may be obtained from The Coca-Cola Company.. Licensed under Fair use of copyrighted material in the context of Coke zero via Wikipedia - "Vanilla cola can" by The logo may be obtained from Vanilla Coke.. Licensed under Fair use of copyrighted material in the context of Vanilla Coke via Wikipedia - "New Coke can". Via Wikipedia - "Coca-Cola lata" by Eduardo Sellan III - Own work. Licensed under Public domain via Wikimedia Commons - "Aspirin1" by User Mosesofmason on zh.wikipedia - Originally from zh.wikipedia; zh:image:aspirin.jpg. Licensed under Creative Commons Attribution-Share Alike 3.0 via Wikimedia Commons - "Arm & Hammer logo" by The logo may be obtained from Arm & Hammer.. Licensed under Fair use of copyrighted material in the context of Arm & Hammer via Wikipedia